[thirdparty/systemd.git] / src / nspawn / nspawn-patch-uid.c

/* SPDX-License-Identifier: LGPL-2.1+ */

#include <fcntl.h>
#include <linux/magic.h>
#if HAVE_ACL
#include <sys/acl.h>
#endif
#include <sys/stat.h>
#include <sys/statvfs.h>
#include <sys/vfs.h>
#include <unistd.h>

#include "acl-util.h"
#include "dirent-util.h"
#include "fd-util.h"
#include "fs-util.h"
#include "missing.h"
#include "nspawn-def.h"
#include "nspawn-patch-uid.h"
#include "stat-util.h"
#include "stdio-util.h"
#include "string-util.h"
#include "strv.h"
#include "user-util.h"

#if HAVE_ACL

static int get_acl(int fd, const char *name, acl_type_t type, acl_t *ret) {
        char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
        acl_t acl;

        assert(fd >= 0);
        assert(ret);

        if (name) {
                _cleanup_close_ int child_fd = -1;

                child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
                if (child_fd < 0)
                        return -errno;

                xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
                acl = acl_get_file(procfs_path, type);
        } else if (type == ACL_TYPE_ACCESS)
                acl = acl_get_fd(fd);
        else {
                xsprintf(procfs_path, "/proc/self/fd/%i", fd);
                acl = acl_get_file(procfs_path, type);
        }
        if (!acl)
                return -errno;

        *ret = acl;
        return 0;
}

static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
        char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
        int r;

        assert(fd >= 0);
        assert(acl);

        if (name) {
                _cleanup_close_ int child_fd = -1;

                child_fd = openat(fd, name, O_PATH|O_CLOEXEC|O_NOFOLLOW);
                if (child_fd < 0)
                        return -errno;

                xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
                r = acl_set_file(procfs_path, type, acl);
        } else if (type == ACL_TYPE_ACCESS)
                r = acl_set_fd(fd, acl);
        else {
                xsprintf(procfs_path, "/proc/self/fd/%i", fd);
                r = acl_set_file(procfs_path, type, acl);
        }
        if (r < 0)
                return -errno;

        return 0;
}

static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
        _cleanup_(acl_freep) acl_t copy = NULL;
        acl_entry_t i;
        int r;

        assert(acl);
        assert(ret);

        r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
        if (r < 0)
                return -errno;
        while (r > 0) {
                uid_t *old_uid, new_uid;
                bool modify = false;
                acl_tag_t tag;

                if (acl_get_tag_type(i, &tag) < 0)
                        return -errno;

                if (IN_SET(tag, ACL_USER, ACL_GROUP)) {

                        /* We don't distuingish here between uid_t and gid_t, let's make sure the compiler checks that
                         * this is actually OK */
                        assert_cc(sizeof(uid_t) == sizeof(gid_t));

                        old_uid = acl_get_qualifier(i);
                        if (!old_uid)
                                return -errno;

                        new_uid = shift | (*old_uid & UINT32_C(0xFFFF));
                        if (!uid_is_valid(new_uid))
                                return -EINVAL;

                        modify = new_uid != *old_uid;
                        if (modify && !copy) {
                                int n;

                                /* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
                                 * beginning, so that we copy all entries, starting from the first, this time. */

                                n = acl_entries(acl);
                                if (n < 0)
                                        return -errno;

                                copy = acl_init(n);
                                if (!copy)
                                        return -errno;

                                /* Seek back to the beginning */
                                r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
                                if (r < 0)
                                        return -errno;
                                continue;
                        }
                }

                if (copy) {
                        acl_entry_t new_entry;

                        if (acl_create_entry(&copy, &new_entry) < 0)
                                return -errno;

                        if (acl_copy_entry(new_entry, i) < 0)
                                return -errno;

                        if (modify)
                                if (acl_set_qualifier(new_entry, &new_uid) < 0)
                                        return -errno;
                }

                r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
                if (r < 0)
                        return -errno;
        }

        *ret = TAKE_PTR(copy);

        return !!*ret;
}

static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
        _cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
        bool changed = false;
        int r;

        assert(fd >= 0);
        assert(st);

        /* ACLs are not supported on symlinks, there's no point in trying */
        if (S_ISLNK(st->st_mode))
                return 0;

        r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
        if (r == -EOPNOTSUPP)
                return 0;
        if (r < 0)
                return r;

        r = shift_acl(acl, shift, &shifted);
        if (r < 0)
                return r;
        if (r > 0) {
                r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
                if (r < 0)
                        return r;

                changed = true;
        }

        if (S_ISDIR(st->st_mode)) {
                acl_free(acl);
                acl_free(shifted);

                acl = shifted = NULL;

                r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
                if (r < 0)
                        return r;

                r = shift_acl(acl, shift, &shifted);
                if (r < 0)
                        return r;
                if (r > 0) {
                        r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
                        if (r < 0)
                                return r;

                        changed = true;
                }
        }

        return changed;
}

#else

static int patch_acls(int fd, const char *name, const struct stat *st, uid_t shift) {
        return 0;
}

#endif

static int patch_fd(int fd, const char *name, const struct stat *st, uid_t shift) {
        uid_t new_uid;
        gid_t new_gid;
        bool changed = false;
        int r;

        assert(fd >= 0);
        assert(st);

        new_uid =         shift | (st->st_uid & UINT32_C(0xFFFF));
        new_gid = (gid_t) shift | (st->st_gid & UINT32_C(0xFFFF));

        if (!uid_is_valid(new_uid) || !gid_is_valid(new_gid))
                return -EINVAL;

        if (st->st_uid != new_uid || st->st_gid != new_gid) {
                if (name)
                        r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
                else
                        r = fchown(fd, new_uid, new_gid);
                if (r < 0)
                        return -errno;

                /* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
                if (name) {
                        if (!S_ISLNK(st->st_mode))
                                r = fchmodat(fd, name, st->st_mode, 0);
                        else /* AT_SYMLINK_NOFOLLOW is not available for fchmodat() */
                                r = 0;
                } else
                        r = fchmod(fd, st->st_mode);
                if (r < 0)
                        return -errno;

                changed = true;
        }

        r = patch_acls(fd, name, st, shift);
        if (r < 0)
                return r;

        return r > 0 || changed;
}

/*
 * Check if the filesystem is fully compatible with user namespaces or
 * UID/GID patching. Some filesystems in this list can be fully mounted inside
 * user namespaces, however their inodes may relate to host resources or only
 * valid in the global user namespace, therefore no patching should be applied.
 */
static int is_fs_fully_userns_compatible(const struct statfs *sfs) {

        assert(sfs);

        return F_TYPE_EQUAL(sfs->f_type, BINFMTFS_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, CGROUP_SUPER_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, CGROUP2_SUPER_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, DEBUGFS_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, DEVPTS_SUPER_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, EFIVARFS_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, HUGETLBFS_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, MQUEUE_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, PROC_SUPER_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, PSTOREFS_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, SELINUX_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, SMACK_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, SECURITYFS_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, BPF_FS_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, TRACEFS_MAGIC) ||
               F_TYPE_EQUAL(sfs->f_type, SYSFS_MAGIC);
}

static int recurse_fd(int fd, bool donate_fd, const struct stat *st, uid_t shift, bool is_toplevel) {
        _cleanup_closedir_ DIR *d = NULL;
        bool changed = false;
        struct statfs sfs;
        int r;

        assert(fd >= 0);

        if (fstatfs(fd, &sfs) < 0)
                return -errno;

        /* We generally want to permit crossing of mount boundaries when patching the UIDs/GIDs. However, we probably
         * shouldn't do this for /proc and /sys if that is already mounted into place. Hence, let's stop the recursion
         * when we hit procfs, sysfs or some other special file systems. */

        r = is_fs_fully_userns_compatible(&sfs);
        if (r < 0)
                goto finish;
        if (r > 0) {
                r = 0; /* don't recurse */
                goto finish;
        }

        /* Also, if we hit a read-only file system, then don't bother, skip the whole subtree */
        if ((sfs.f_flags & ST_RDONLY) ||
            access_fd(fd, W_OK) == -EROFS)
                goto read_only;

        if (S_ISDIR(st->st_mode)) {
                struct dirent *de;

                if (!donate_fd) {
                        int copy;

                        copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
                        if (copy < 0) {
                                r = -errno;
                                goto finish;
                        }

                        fd = copy;
                        donate_fd = true;
                }

                d = fdopendir(fd);
                if (!d) {
                        r = -errno;
                        goto finish;
                }
                fd = -1;

                FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
                        struct stat fst;

                        if (dot_or_dot_dot(de->d_name))
                                continue;

                        if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
                                r = -errno;
                                goto finish;
                        }

                        if (S_ISDIR(fst.st_mode)) {
                                int subdir_fd;

                                subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
                                if (subdir_fd < 0) {
                                        r = -errno;
                                        goto finish;

                                }

                                r = recurse_fd(subdir_fd, true, &fst, shift, false);
                                if (r < 0)
                                        goto finish;
                                if (r > 0)
                                        changed = true;

                        } else {
                                r = patch_fd(dirfd(d), de->d_name, &fst, shift);
                                if (r < 0)
                                        goto finish;
                                if (r > 0)
                                        changed = true;
                        }
                }
        }

        /* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
         * directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
         * tree is properly chown()ed already. */
        r = patch_fd(d ? dirfd(d) : fd, NULL, st, shift);
        if (r == -EROFS)
                goto read_only;
        if (r > 0)
                changed = true;

        r = changed;
        goto finish;

read_only:
        if (!is_toplevel) {
                _cleanup_free_ char *name = NULL;

                /* When we hit a ready-only subtree we simply skip it, but log about it. */
                (void) fd_get_path(fd, &name);
                log_debug("Skippping read-only file or directory %s.", strna(name));
                r = changed;
        }

finish:
        if (donate_fd)
                safe_close(fd);

        return r;
}

static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
        struct stat st;
        int r;

        assert(fd >= 0);

        /* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
         * OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
         * following the concept that the upper 16bit of a UID identify the container, and the lower 16bit are the actual
         * UID within the container. */

        if ((shift & 0xFFFF) != 0) {
                /* We only support containers where the shift starts at a 2^16 boundary */
                r = -EOPNOTSUPP;
                goto finish;
        }

        if (shift == UID_BUSY_BASE) {
                r = -EINVAL;
                goto finish;
        }

        if (range != 0x10000) {
                /* We only support containers with 16bit UID ranges for the patching logic */
                r = -EOPNOTSUPP;
                goto finish;
        }

        if (fstat(fd, &st) < 0) {
                r = -errno;
                goto finish;
        }

        if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
                /* We only support containers where the uid/gid container ID match */
                r = -EBADE;
                goto finish;
        }

        /* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
         * that if the top-level dir has the right upper 16bit assigned, then everything below will have too... */
        if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
                return 0;

        /* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
         * range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
         * chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */

        if ((st.st_uid & UID_BUSY_MASK) != UID_BUSY_BASE) {
                if (fchown(fd,
                           UID_BUSY_BASE | (st.st_uid & ~UID_BUSY_MASK),
                           (gid_t) UID_BUSY_BASE | (st.st_gid & ~(gid_t) UID_BUSY_MASK)) < 0) {
                        r = -errno;
                        goto finish;
                }
        }

        return recurse_fd(fd, donate_fd, &st, shift, true);

finish:
        if (donate_fd)
                safe_close(fd);

        return r;
}

int path_patch_uid(const char *path, uid_t shift, uid_t range) {
        int fd;

        fd = open(path, O_RDONLY|O_NONBLOCK|O_DIRECTORY|O_CLOEXEC|O_NOFOLLOW|O_NOATIME);
        if (fd < 0)
                return -errno;

        return fd_patch_uid_internal(fd, true, shift, range);
}
Commit	Line	Data
53e1b683	1	/* SPDX-License-Identifier: LGPL-2.1+ */
7336138e LP	2
7336138e LP	3	#include <fcntl.h>
88cd066e	4	#include <linux/magic.h>
349cc4a5	5	#if HAVE_ACL
7336138e LP	6	#include <sys/acl.h>
	7	#endif
	8	#include <sys/stat.h>
3603efde	9	#include <sys/statvfs.h>
88cd066e	10	#include <sys/vfs.h>
7336138e LP	11	#include <unistd.h>
	12
	13	#include "acl-util.h"
	14	#include "dirent-util.h"
	15	#include "fd-util.h"
3603efde	16	#include "fs-util.h"
88cd066e	17	#include "missing.h"
3603efde	18	#include "nspawn-def.h"
7336138e	19	#include "nspawn-patch-uid.h"
88cd066e	20	#include "stat-util.h"
7336138e LP	21	#include "stdio-util.h"
	22	#include "string-util.h"
	23	#include "strv.h"
	24	#include "user-util.h"
	25
349cc4a5	26	#if HAVE_ACL
7336138e LP	27
7336138e LP	28	static int get_acl(int fd, const char name, acl_type_t type, acl_t ret) {
fbd0b64f	29	char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
7336138e LP	30	acl_t acl;
	31
	32	assert(fd >= 0);
	33	assert(ret);
	34
	35	if (name) {
	36	_cleanup_close_ int child_fd = -1;
	37
	38	child_fd = openat(fd, name, O_PATH\|O_CLOEXEC\|O_NOFOLLOW);
	39	if (child_fd < 0)
	40	return -errno;
	41
	42	xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
	43	acl = acl_get_file(procfs_path, type);
	44	} else if (type == ACL_TYPE_ACCESS)
	45	acl = acl_get_fd(fd);
	46	else {
	47	xsprintf(procfs_path, "/proc/self/fd/%i", fd);
	48	acl = acl_get_file(procfs_path, type);
	49	}
	50	if (!acl)
	51	return -errno;
	52
	53	*ret = acl;
	54	return 0;
	55	}
	56
	57	static int set_acl(int fd, const char *name, acl_type_t type, acl_t acl) {
fbd0b64f	58	char procfs_path[STRLEN("/proc/self/fd/") + DECIMAL_STR_MAX(int) + 1];
7336138e LP	59	int r;
	60
	61	assert(fd >= 0);
	62	assert(acl);
	63
	64	if (name) {
	65	_cleanup_close_ int child_fd = -1;
	66
	67	child_fd = openat(fd, name, O_PATH\|O_CLOEXEC\|O_NOFOLLOW);
	68	if (child_fd < 0)
	69	return -errno;
	70
	71	xsprintf(procfs_path, "/proc/self/fd/%i", child_fd);
	72	r = acl_set_file(procfs_path, type, acl);
	73	} else if (type == ACL_TYPE_ACCESS)
	74	r = acl_set_fd(fd, acl);
	75	else {
	76	xsprintf(procfs_path, "/proc/self/fd/%i", fd);
	77	r = acl_set_file(procfs_path, type, acl);
	78	}
	79	if (r < 0)
	80	return -errno;
	81
	82	return 0;
	83	}
	84
	85	static int shift_acl(acl_t acl, uid_t shift, acl_t *ret) {
	86	_cleanup_(acl_freep) acl_t copy = NULL;
	87	acl_entry_t i;
	88	int r;
	89
	90	assert(acl);
	91	assert(ret);
	92
	93	r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
	94	if (r < 0)
	95	return -errno;
	96	while (r > 0) {
	97	uid_t *old_uid, new_uid;
	98	bool modify = false;
	99	acl_tag_t tag;
	100
	101	if (acl_get_tag_type(i, &tag) < 0)
	102	return -errno;
	103
	104	if (IN_SET(tag, ACL_USER, ACL_GROUP)) {
	105
	106	/* We don't distuingish here between uid_t and gid_t, let's make sure the compiler checks that
	107	* this is actually OK */
	108	assert_cc(sizeof(uid_t) == sizeof(gid_t));
	109
	110	old_uid = acl_get_qualifier(i);
	111	if (!old_uid)
	112	return -errno;
	113
	114	new_uid = shift \| (*old_uid & UINT32_C(0xFFFF));
	115	if (!uid_is_valid(new_uid))
	116	return -EINVAL;
	117
	118	modify = new_uid != *old_uid;
	119	if (modify && !copy) {
	120	int n;
	121
	122	/* There's no copy of the ACL yet? if so, let's create one, and start the loop from the
123	* beginning, so that we copy all entries, starting from the first, this time. */
124
125	n = acl_entries(acl);
126	if (n < 0)
127	return -errno;
128
129	copy = acl_init(n);
130	if (!copy)
131	return -errno;
132
133	/* Seek back to the beginning */
134	r = acl_get_entry(acl, ACL_FIRST_ENTRY, &i);
135	if (r < 0)
136	return -errno;
137	continue;
138	}
139	}
140
141	if (copy) {
142	acl_entry_t new_entry;
143
144	if (acl_create_entry(&copy, &new_entry) < 0)
145	return -errno;
146
147	if (acl_copy_entry(new_entry, i) < 0)
148	return -errno;
149
150	if (modify)
151	if (acl_set_qualifier(new_entry, &new_uid) < 0)
152	return -errno;
153	}
154
155	r = acl_get_entry(acl, ACL_NEXT_ENTRY, &i);
156	if (r < 0)
157	return -errno;
158	}
159
1cc6c93a	160	*ret = TAKE_PTR(copy);
7336138e LP	161
	162	return !!*ret;
	163	}
	164
	165	static int patch_acls(int fd, const char name, const struct stat st, uid_t shift) {
	166	_cleanup_(acl_freep) acl_t acl = NULL, shifted = NULL;
	167	bool changed = false;
	168	int r;
	169
	170	assert(fd >= 0);
	171	assert(st);
	172
	173	/* ACLs are not supported on symlinks, there's no point in trying */
	174	if (S_ISLNK(st->st_mode))
	175	return 0;
	176
	177	r = get_acl(fd, name, ACL_TYPE_ACCESS, &acl);
	178	if (r == -EOPNOTSUPP)
	179	return 0;
	180	if (r < 0)
	181	return r;
	182
	183	r = shift_acl(acl, shift, &shifted);
	184	if (r < 0)
	185	return r;
	186	if (r > 0) {
	187	r = set_acl(fd, name, ACL_TYPE_ACCESS, shifted);
	188	if (r < 0)
	189	return r;
	190
	191	changed = true;
	192	}
	193
	194	if (S_ISDIR(st->st_mode)) {
	195	acl_free(acl);
	196	acl_free(shifted);
	197
	198	acl = shifted = NULL;
	199
	200	r = get_acl(fd, name, ACL_TYPE_DEFAULT, &acl);
	201	if (r < 0)
	202	return r;
	203
	204	r = shift_acl(acl, shift, &shifted);
	205	if (r < 0)
	206	return r;
	207	if (r > 0) {
	208	r = set_acl(fd, name, ACL_TYPE_DEFAULT, shifted);
	209	if (r < 0)
	210	return r;
	211
	212	changed = true;
	213	}
	214	}
	215
	216	return changed;
	217	}
	218
	219	#else
	220
	221	static int patch_acls(int fd, const char name, const struct stat st, uid_t shift) {
	222	return 0;
	223	}
	224
225	#endif
226
227	static int patch_fd(int fd, const char name, const struct stat st, uid_t shift) {
228	uid_t new_uid;
229	gid_t new_gid;
230	bool changed = false;
231	int r;
232
233	assert(fd >= 0);
234	assert(st);
235
236	new_uid = shift \| (st->st_uid & UINT32_C(0xFFFF));
237	new_gid = (gid_t) shift \| (st->st_gid & UINT32_C(0xFFFF));
238
239	if (!uid_is_valid(new_uid) \|\| !gid_is_valid(new_gid))
240	return -EINVAL;
241
242	if (st->st_uid != new_uid \|\| st->st_gid != new_gid) {
243	if (name)
244	r = fchownat(fd, name, new_uid, new_gid, AT_SYMLINK_NOFOLLOW);
245	else
246	r = fchown(fd, new_uid, new_gid);
247	if (r < 0)
248	return -errno;
249
250	/* The Linux kernel alters the mode in some cases of chown(). Let's undo this. */
0c6aeb46 LP	251	if (name) {
	252	if (!S_ISLNK(st->st_mode))
	253	r = fchmodat(fd, name, st->st_mode, 0);
	254	else /* AT_SYMLINK_NOFOLLOW is not available for fchmodat() */
	255	r = 0;
	256	} else
7336138e LP	257	r = fchmod(fd, st->st_mode);
	258	if (r < 0)
	259	return -errno;
	260
	261	changed = true;
	262	}
	263
	264	r = patch_acls(fd, name, st, shift);
	265	if (r < 0)
	266	return r;
	267
	268	return r > 0 \|\| changed;
	269	}
	270
231bfb1b DH	271	/*
	272	* Check if the filesystem is fully compatible with user namespaces or
	273	* UID/GID patching. Some filesystems in this list can be fully mounted inside
	274	* user namespaces, however their inodes may relate to host resources or only
	275	* valid in the global user namespace, therefore no patching should be applied.
	276	*/
3603efde LP	277	static int is_fs_fully_userns_compatible(const struct statfs *sfs) {
	278
	279	assert(sfs);
	280
	281	return F_TYPE_EQUAL(sfs->f_type, BINFMTFS_MAGIC) \|\|
	282	F_TYPE_EQUAL(sfs->f_type, CGROUP_SUPER_MAGIC) \|\|
	283	F_TYPE_EQUAL(sfs->f_type, CGROUP2_SUPER_MAGIC) \|\|
	284	F_TYPE_EQUAL(sfs->f_type, DEBUGFS_MAGIC) \|\|
	285	F_TYPE_EQUAL(sfs->f_type, DEVPTS_SUPER_MAGIC) \|\|
	286	F_TYPE_EQUAL(sfs->f_type, EFIVARFS_MAGIC) \|\|
	287	F_TYPE_EQUAL(sfs->f_type, HUGETLBFS_MAGIC) \|\|
	288	F_TYPE_EQUAL(sfs->f_type, MQUEUE_MAGIC) \|\|
	289	F_TYPE_EQUAL(sfs->f_type, PROC_SUPER_MAGIC) \|\|
	290	F_TYPE_EQUAL(sfs->f_type, PSTOREFS_MAGIC) \|\|
	291	F_TYPE_EQUAL(sfs->f_type, SELINUX_MAGIC) \|\|
	292	F_TYPE_EQUAL(sfs->f_type, SMACK_MAGIC) \|\|
	293	F_TYPE_EQUAL(sfs->f_type, SECURITYFS_MAGIC) \|\|
	294	F_TYPE_EQUAL(sfs->f_type, BPF_FS_MAGIC) \|\|
	295	F_TYPE_EQUAL(sfs->f_type, TRACEFS_MAGIC) \|\|
	296	F_TYPE_EQUAL(sfs->f_type, SYSFS_MAGIC);
88cd066e LP	297	}
88cd066e LP	298
4aeb20f5	299	static int recurse_fd(int fd, bool donate_fd, const struct stat *st, uid_t shift, bool is_toplevel) {
3603efde	300	_cleanup_closedir_ DIR *d = NULL;
7336138e	301	bool changed = false;
3603efde	302	struct statfs sfs;
7336138e LP	303	int r;
	304
	305	assert(fd >= 0);
	306
3603efde LP	307	if (fstatfs(fd, &sfs) < 0)
	308	return -errno;
	309
	310	/* We generally want to permit crossing of mount boundaries when patching the UIDs/GIDs. However, we probably
	311	* shouldn't do this for /proc and /sys if that is already mounted into place. Hence, let's stop the recursion
	312	* when we hit procfs, sysfs or some other special file systems. */
	313
	314	r = is_fs_fully_userns_compatible(&sfs);
88cd066e LP	315	if (r < 0)
	316	goto finish;
	317	if (r > 0) {
	318	r = 0; /* don't recurse */
	319	goto finish;
	320	}
	321
3603efde LP	322	/* Also, if we hit a read-only file system, then don't bother, skip the whole subtree */
	323	if ((sfs.f_flags & ST_RDONLY) \|\|
	324	access_fd(fd, W_OK) == -EROFS)
	325	goto read_only;
7336138e LP	326
7336138e LP	327	if (S_ISDIR(st->st_mode)) {
7336138e LP	328	struct dirent *de;
	329
	330	if (!donate_fd) {
	331	int copy;
	332
	333	copy = fcntl(fd, F_DUPFD_CLOEXEC, 3);
88cd066e LP	334	if (copy < 0) {
	335	r = -errno;
	336	goto finish;
	337	}
7336138e LP	338
	339	fd = copy;
	340	donate_fd = true;
	341	}
	342
	343	d = fdopendir(fd);
	344	if (!d) {
	345	r = -errno;
	346	goto finish;
	347	}
	348	fd = -1;
	349
	350	FOREACH_DIRENT_ALL(de, d, r = -errno; goto finish) {
	351	struct stat fst;
	352
49bfc877	353	if (dot_or_dot_dot(de->d_name))
7336138e LP	354	continue;
	355
	356	if (fstatat(dirfd(d), de->d_name, &fst, AT_SYMLINK_NOFOLLOW) < 0) {
	357	r = -errno;
	358	goto finish;
	359	}
	360
	361	if (S_ISDIR(fst.st_mode)) {
	362	int subdir_fd;
	363
	364	subdir_fd = openat(dirfd(d), de->d_name, O_RDONLY\|O_NONBLOCK\|O_DIRECTORY\|O_CLOEXEC\|O_NOFOLLOW\|O_NOATIME);
	365	if (subdir_fd < 0) {
	366	r = -errno;
	367	goto finish;
	368
	369	}
	370
4aeb20f5	371	r = recurse_fd(subdir_fd, true, &fst, shift, false);
7336138e LP	372	if (r < 0)
	373	goto finish;
	374	if (r > 0)
	375	changed = true;
	376
	377	} else {
	378	r = patch_fd(dirfd(d), de->d_name, &fst, shift);
	379	if (r < 0)
	380	goto finish;
	381	if (r > 0)
	382	changed = true;
	383	}
	384	}
	385	}
	386
3603efde LP	387	/* After we descended, also patch the directory itself. It's key to do this in this order so that the top-level
	388	* directory is patched as very last object in the tree, so that we can use it as quick indicator whether the
	389	* tree is properly chown()ed already. */
	390	r = patch_fd(d ? dirfd(d) : fd, NULL, st, shift);
	391	if (r == -EROFS)
	392	goto read_only;
	393	if (r > 0)
	394	changed = true;
	395
7336138e	396	r = changed;
3603efde LP	397	goto finish;
	398
	399	read_only:
	400	if (!is_toplevel) {
	401	_cleanup_free_ char *name = NULL;
	402
	403	/* When we hit a ready-only subtree we simply skip it, but log about it. */
	404	(void) fd_get_path(fd, &name);
	405	log_debug("Skippping read-only file or directory %s.", strna(name));
	406	r = changed;
	407	}
7336138e LP	408
	409	finish:
	410	if (donate_fd)
	411	safe_close(fd);
	412
	413	return r;
	414	}
	415
	416	static int fd_patch_uid_internal(int fd, bool donate_fd, uid_t shift, uid_t range) {
	417	struct stat st;
	418	int r;
	419
	420	assert(fd >= 0);
	421
	422	/* Recursively adjusts the UID/GIDs of all files of a directory tree. This is used to automatically fix up an
	423	* OS tree to the used user namespace UID range. Note that this automatic adjustment only works for UID ranges
	424	* following the concept that the upper 16bit of a UID identify the container, and the lower 16bit are the actual
	425	* UID within the container. */
	426
	427	if ((shift & 0xFFFF) != 0) {
	428	/* We only support containers where the shift starts at a 2^16 boundary */
	429	r = -EOPNOTSUPP;
	430	goto finish;
	431	}
	432
3603efde LP	433	if (shift == UID_BUSY_BASE) {
	434	r = -EINVAL;
	435	goto finish;
	436	}
	437
7336138e LP	438	if (range != 0x10000) {
	439	/* We only support containers with 16bit UID ranges for the patching logic */
	440	r = -EOPNOTSUPP;
	441	goto finish;
	442	}
	443
	444	if (fstat(fd, &st) < 0) {
	445	r = -errno;
	446	goto finish;
	447	}
	448
	449	if ((uint32_t) st.st_uid >> 16 != (uint32_t) st.st_gid >> 16) {
	450	/* We only support containers where the uid/gid container ID match */
	451	r = -EBADE;
	452	goto finish;
	453	}
	454
	455	/* Try to detect if the range is already right. Of course, this a pretty drastic optimization, as we assume
	456	* that if the top-level dir has the right upper 16bit assigned, then everything below will have too... */
	457	if (((uint32_t) (st.st_uid ^ shift) >> 16) == 0)
	458	return 0;
	459
3603efde LP	460	/* Before we start recursively chowning, mark the top-level dir as "busy" by chowning it to the "busy"
	461	* range. Should we be interrupted in the middle of our work, we'll see it owned by this user and will start
	462	* chown()ing it again, unconditionally, as the busy UID is not a valid UID we'd everpick for ourselves. */
	463
	464	if ((st.st_uid & UID_BUSY_MASK) != UID_BUSY_BASE) {
	465	if (fchown(fd,
	466	UID_BUSY_BASE \| (st.st_uid & ~UID_BUSY_MASK),
	467	(gid_t) UID_BUSY_BASE \| (st.st_gid & ~(gid_t) UID_BUSY_MASK)) < 0) {
	468	r = -errno;
	469	goto finish;
	470	}
	471	}
	472
4aeb20f5	473	return recurse_fd(fd, donate_fd, &st, shift, true);
7336138e LP	474
	475	finish:
	476	if (donate_fd)
	477	safe_close(fd);
	478
	479	return r;
	480	}
	481
7336138e LP	482	int path_patch_uid(const char *path, uid_t shift, uid_t range) {
	483	int fd;
	484
	485	fd = open(path, O_RDONLY\|O_NONBLOCK\|O_DIRECTORY\|O_CLOEXEC\|O_NOFOLLOW\|O_NOATIME);
	486	if (fd < 0)
	487	return -errno;
	488
	489	return fd_patch_uid_internal(fd, true, shift, range);
	490	}